AI Guides › Skills & Agents

The Skill You Write Once And The Prompt You Write Every Time

By Nigel Guy · 3 min read

The usual way people decide between a skill and a repeated prompt is frequency: do this often enough and eventually you build a skill for it. That's a reasonable rough guide, but it answers the wrong question. The question that actually decides it is what varies between the times you do the task — because a task you do daily but where only the wording changes is still, correctly, just a prompt, and a task you've only done twice can already need a skill if getting the sequence right matters more than getting the phrasing right.

The mechanism

1. Pull up the last three times you did this task. Not from memory — actually look at what you asked for, if you can, or reconstruct it as honestly as you can.
2. Note what actually changed between them. Was it just the inputs — the name, the date, the numbers — with the same steps happening in the same order underneath? Or did the sequence of steps itself change, branch, or get skipped depending on the situation?
3. If only the inputs changed, a template with blanks is enough. Save the version of the prompt that worked, mark where the variables go, and stop there — building more structure than this earns you nothing.
4. If the steps need to hold even when you're rushed or careless, that's the actual signal for a skill. The value a skill adds over a prompt isn't wording — it's that the steps happen in order and the guardrails apply regardless of how much attention you're paying that particular day. A prompt you retype from memory doesn't protect against skipping a step when you're in a hurry; a skill's structure does.
5. Let the task graduate later if its nature changes. A task that started as pure input-swapping can grow real branching complexity over time — that's the moment to revisit, not a sign you decided wrong originally.

What to skip

Skip building a skill pre-emptively for a task you've only done once — you don't yet know what varies between instances, and guessing at the guardrails before you have real cases produces a skill built around imagined failure modes instead of actual ones. And skip re-typing a full explanation from scratch each time for a task where genuinely only the inputs differ — that's exactly the case the saved-prompt template exists to solve, and building a skill for it adds ceremony without adding reliability.

Guardrails

• This variance test is a different lens on the same underlying question covered elsewhere in this library as skill-versus-saved-prompt — stakes and cost of getting it wrong matter too; use whichever framing actually clarifies your specific case.
• Getting this judgement wrong in either direction is recoverable — an over-built skill can be simplified back down, and an under-built prompt can be upgraded once it's clearly earned it. Neither mistake is expensive to fix, which is a reason not to agonise over the initial call.
• A prompt template still needs periodic review just like a skill does — a template built against last year's tool behaviour can quietly produce worse results without any structural change alerting you.

All 751 AI guides · JulieMango plans from £17/mo